[linux.git] / kernel / latencytop.c

// SPDX-License-Identifier: GPL-2.0-only
/*
 * latencytop.c: Latency display infrastructure
 *
 * (C) Copyright 2008 Intel Corporation
 * Author: Arjan van de Ven <[email protected]>
 */

/*
 * CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
 * used by the "latencytop" userspace tool. The latency that is tracked is not
 * the 'traditional' interrupt latency (which is primarily caused by something
 * else consuming CPU), but instead, it is the latency an application encounters
 * because the kernel sleeps on its behalf for various reasons.
 *
 * This code tracks 2 levels of statistics:
 * 1) System level latency
 * 2) Per process latency
 *
 * The latency is stored in fixed sized data structures in an accumulated form;
 * if the "same" latency cause is hit twice, this will be tracked as one entry
 * in the data structure. Both the count, total accumulated latency and maximum
 * latency are tracked in this data structure. When the fixed size structure is
 * full, no new causes are tracked until the buffer is flushed by writing to
 * the /proc file; the userspace tool does this on a regular basis.
 *
 * A latency cause is identified by a stringified backtrace at the point that
 * the scheduler gets invoked. The userland tool will use this string to
 * identify the cause of the latency in human readable form.
 *
 * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
 * These files look like this:
 *
 * Latency Top version : v0.1
 * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
 * |    |    |    |
 * |    |    |    +----> the stringified backtrace
 * |    |    +---------> The maximum latency for this entry in microseconds
 * |    +--------------> The accumulated latency for this entry (microseconds)
 * +-------------------> The number of times this entry is hit
 *
 * (note: the average latency is the accumulated latency divided by the number
 * of times)
 */

#include <linux/kallsyms.h>
#include <linux/seq_file.h>
#include <linux/notifier.h>
#include <linux/spinlock.h>
#include <linux/proc_fs.h>
#include <linux/latencytop.h>
#include <linux/export.h>
#include <linux/sched.h>
#include <linux/sched/debug.h>
#include <linux/sched/stat.h>
#include <linux/list.h>
#include <linux/stacktrace.h>
#include <linux/sysctl.h>

static DEFINE_RAW_SPINLOCK(latency_lock);

#define MAXLR 128
static struct latency_record latency_record[MAXLR];

int latencytop_enabled;

#ifdef CONFIG_SYSCTL
static int sysctl_latencytop(struct ctl_table *table, int write, void *buffer,
		size_t *lenp, loff_t *ppos)
{
	int err;

	err = proc_dointvec(table, write, buffer, lenp, ppos);
	if (latencytop_enabled)
		force_schedstat_enabled();

	return err;
}

static struct ctl_table latencytop_sysctl[] = {
	{
		.procname   = "latencytop",
		.data       = &latencytop_enabled,
		.maxlen     = sizeof(int),
		.mode       = 0644,
		.proc_handler   = sysctl_latencytop,
	},
	{}
};
#endif

void clear_tsk_latency_tracing(struct task_struct *p)
{
	unsigned long flags;

	raw_spin_lock_irqsave(&latency_lock, flags);
	memset(&p->latency_record, 0, sizeof(p->latency_record));
	p->latency_record_count = 0;
	raw_spin_unlock_irqrestore(&latency_lock, flags);
}

static void clear_global_latency_tracing(void)
{
	unsigned long flags;

	raw_spin_lock_irqsave(&latency_lock, flags);
	memset(&latency_record, 0, sizeof(latency_record));
	raw_spin_unlock_irqrestore(&latency_lock, flags);
}

static void __sched
account_global_scheduler_latency(struct task_struct *tsk,
				 struct latency_record *lat)
{
	int firstnonnull = MAXLR;
	int i;

	/* skip kernel threads for now */
	if (!tsk->mm)
		return;

	for (i = 0; i < MAXLR; i++) {
		int q, same = 1;

		/* Nothing stored: */
		if (!latency_record[i].backtrace[0]) {
			if (firstnonnull > i)
				firstnonnull = i;
			continue;
		}
		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
			unsigned long record = lat->backtrace[q];

			if (latency_record[i].backtrace[q] != record) {
				same = 0;
				break;
			}

			/* 0 entry marks end of backtrace: */
			if (!record)
				break;
		}
		if (same) {
			latency_record[i].count++;
			latency_record[i].time += lat->time;
			if (lat->time > latency_record[i].max)
				latency_record[i].max = lat->time;
			return;
		}
	}

	i = firstnonnull;
	if (i >= MAXLR)
		return;

	/* Allocted a new one: */
	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
}

/**
 * __account_scheduler_latency - record an occurred latency
 * @tsk - the task struct of the task hitting the latency
 * @usecs - the duration of the latency in microseconds
 * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
 *
 * This function is the main entry point for recording latency entries
 * as called by the scheduler.
 *
 * This function has a few special cases to deal with normal 'non-latency'
 * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
 * since this usually is caused by waiting for events via select() and co.
 *
 * Negative latencies (caused by time going backwards) are also explicitly
 * skipped.
 */
void __sched
__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
{
	unsigned long flags;
	int i, q;
	struct latency_record lat;

	/* Long interruptible waits are generally user requested... */
	if (inter && usecs > 5000)
		return;

	/* Negative sleeps are time going backwards */
	/* Zero-time sleeps are non-interesting */
	if (usecs <= 0)
		return;

	memset(&lat, 0, sizeof(lat));
	lat.count = 1;
	lat.time = usecs;
	lat.max = usecs;

	stack_trace_save_tsk(tsk, lat.backtrace, LT_BACKTRACEDEPTH, 0);

	raw_spin_lock_irqsave(&latency_lock, flags);

	account_global_scheduler_latency(tsk, &lat);

	for (i = 0; i < tsk->latency_record_count; i++) {
		struct latency_record *mylat;
		int same = 1;

		mylat = &tsk->latency_record[i];
		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
			unsigned long record = lat.backtrace[q];

			if (mylat->backtrace[q] != record) {
				same = 0;
				break;
			}

			/* 0 entry is end of backtrace */
			if (!record)
				break;
		}
		if (same) {
			mylat->count++;
			mylat->time += lat.time;
			if (lat.time > mylat->max)
				mylat->max = lat.time;
			goto out_unlock;
		}
	}

	/*
	 * short term hack; if we're > 32 we stop; future we recycle:
	 */
	if (tsk->latency_record_count >= LT_SAVECOUNT)
		goto out_unlock;

	/* Allocated a new one: */
	i = tsk->latency_record_count++;
	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));

out_unlock:
	raw_spin_unlock_irqrestore(&latency_lock, flags);
}

static int lstats_show(struct seq_file *m, void *v)
{
	int i;

	seq_puts(m, "Latency Top version : v0.1\n");

	for (i = 0; i < MAXLR; i++) {
		struct latency_record *lr = &latency_record[i];

		if (lr->backtrace[0]) {
			int q;
			seq_printf(m, "%i %lu %lu",
				   lr->count, lr->time, lr->max);
			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
				unsigned long bt = lr->backtrace[q];

				if (!bt)
					break;

				seq_printf(m, " %ps", (void *)bt);
			}
			seq_puts(m, "\n");
		}
	}
	return 0;
}

static ssize_t
lstats_write(struct file *file, const char __user *buf, size_t count,
	     loff_t *offs)
{
	clear_global_latency_tracing();

	return count;
}

static int lstats_open(struct inode *inode, struct file *filp)
{
	return single_open(filp, lstats_show, NULL);
}

static const struct proc_ops lstats_proc_ops = {
	.proc_open	= lstats_open,
	.proc_read	= seq_read,
	.proc_write	= lstats_write,
	.proc_lseek	= seq_lseek,
	.proc_release	= single_release,
};

static int __init init_lstats_procfs(void)
{
	proc_create("latency_stats", 0644, NULL, &lstats_proc_ops);
#ifdef CONFIG_SYSCTL
	register_sysctl_init("kernel", latencytop_sysctl);
#endif
	return 0;
}
device_initcall(init_lstats_procfs);
Commit	Line	Data
b886d83c	1	// SPDX-License-Identifier: GPL-2.0-only
9745512c AV	2	/*
	3	* latencytop.c: Latency display infrastructure
	4	*
	5	* (C) Copyright 2008 Intel Corporation
	6	* Author: Arjan van de Ven <[email protected]>
9745512c	7	*/
ad0b0fd5 AV	8
	9	/*
	10	* CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
	11	* used by the "latencytop" userspace tool. The latency that is tracked is not
	12	* the 'traditional' interrupt latency (which is primarily caused by something
	13	* else consuming CPU), but instead, it is the latency an application encounters
	14	* because the kernel sleeps on its behalf for various reasons.
	15	*
	16	* This code tracks 2 levels of statistics:
	17	* 1) System level latency
	18	* 2) Per process latency
	19	*
	20	* The latency is stored in fixed sized data structures in an accumulated form;
	21	* if the "same" latency cause is hit twice, this will be tracked as one entry
	22	* in the data structure. Both the count, total accumulated latency and maximum
	23	* latency are tracked in this data structure. When the fixed size structure is
	24	* full, no new causes are tracked until the buffer is flushed by writing to
	25	* the /proc file; the userspace tool does this on a regular basis.
	26	*
	27	* A latency cause is identified by a stringified backtrace at the point that
	28	* the scheduler gets invoked. The userland tool will use this string to
	29	* identify the cause of the latency in human readable form.
	30	*
	31	* The information is exported via /proc/latency_stats and /proc/<pid>/latency.
	32	* These files look like this:
	33	*
	34	* Latency Top version : v0.1
	35	* 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
	36	* \| \| \| \|
	37	* \| \| \| +----> the stringified backtrace
	38	* \| \| +---------> The maximum latency for this entry in microseconds
	39	* \| +--------------> The accumulated latency for this entry (microseconds)
	40	* +-------------------> The number of times this entry is hit
	41	*
	42	* (note: the average latency is the accumulated latency divided by the number
	43	* of times)
	44	*/
	45
9745512c AV	46	#include <linux/kallsyms.h>
	47	#include <linux/seq_file.h>
	48	#include <linux/notifier.h>
	49	#include <linux/spinlock.h>
	50	#include <linux/proc_fs.h>
cb251765	51	#include <linux/latencytop.h>
9984de1a	52	#include <linux/export.h>
9745512c	53	#include <linux/sched.h>
b17b0153	54	#include <linux/sched/debug.h>
3905f9ad	55	#include <linux/sched/stat.h>
9745512c	56	#include <linux/list.h>
9745512c	57	#include <linux/stacktrace.h>
988f11e0	58	#include <linux/sysctl.h>
9745512c	59
757455d4	60	static DEFINE_RAW_SPINLOCK(latency_lock);
9745512c AV	61
	62	#define MAXLR 128
	63	static struct latency_record latency_record[MAXLR];
	64
	65	int latencytop_enabled;
	66
988f11e0	67	#ifdef CONFIG_SYSCTL
	68	static int sysctl_latencytop(struct ctl_table table, int write, void buffer,
	69	size_t lenp, loff_t ppos)
	70	{
	71	int err;
	72
	73	err = proc_dointvec(table, write, buffer, lenp, ppos);
	74	if (latencytop_enabled)
	75	force_schedstat_enabled();
	76
	77	return err;
	78	}
	79
	80	static struct ctl_table latencytop_sysctl[] = {
	81	{
	82	.procname = "latencytop",
	83	.data = &latencytop_enabled,
	84	.maxlen = sizeof(int),
	85	.mode = 0644,
	86	.proc_handler = sysctl_latencytop,
	87	},
	88	{}
	89	};
	90	#endif
	91
e02c9b0d	92	void clear_tsk_latency_tracing(struct task_struct *p)
9745512c AV	93	{
	94	unsigned long flags;
	95
757455d4	96	raw_spin_lock_irqsave(&latency_lock, flags);
9745512c AV	97	memset(&p->latency_record, 0, sizeof(p->latency_record));
9745512c AV	98	p->latency_record_count = 0;
757455d4	99	raw_spin_unlock_irqrestore(&latency_lock, flags);
9745512c AV	100	}
	101
	102	static void clear_global_latency_tracing(void)
	103	{
	104	unsigned long flags;
	105
757455d4	106	raw_spin_lock_irqsave(&latency_lock, flags);
9745512c	107	memset(&latency_record, 0, sizeof(latency_record));
757455d4	108	raw_spin_unlock_irqrestore(&latency_lock, flags);
9745512c AV	109	}
	110
	111	static void __sched
eaa1809b FF	112	account_global_scheduler_latency(struct task_struct *tsk,
eaa1809b FF	113	struct latency_record *lat)
9745512c	114	{
b8147511	115	int firstnonnull = MAXLR;
9745512c AV	116	int i;
9745512c AV	117
9745512c AV	118	/* skip kernel threads for now */
	119	if (!tsk->mm)
	120	return;
	121
	122	for (i = 0; i < MAXLR; i++) {
19fb518c DA	123	int q, same = 1;
19fb518c DA	124
9745512c AV	125	/* Nothing stored: */
	126	if (!latency_record[i].backtrace[0]) {
	127	if (firstnonnull > i)
	128	firstnonnull = i;
	129	continue;
	130	}
ad0b0fd5	131	for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
19fb518c DA	132	unsigned long record = lat->backtrace[q];
	133
	134	if (latency_record[i].backtrace[q] != record) {
9745512c	135	same = 0;
9745512c	136	break;
19fb518c DA	137	}
19fb518c DA	138
accddc41 TG	139	/* 0 entry marks end of backtrace: */
accddc41 TG	140	if (!record)
9745512c AV	141	break;
	142	}
	143	if (same) {
	144	latency_record[i].count++;
	145	latency_record[i].time += lat->time;
	146	if (lat->time > latency_record[i].max)
	147	latency_record[i].max = lat->time;
	148	return;
	149	}
	150	}
	151
	152	i = firstnonnull;
b8147511	153	if (i >= MAXLR)
9745512c AV	154	return;
	155
	156	/* Allocted a new one: */
	157	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
	158	}
	159
ad0b0fd5	160	/**
25985edc	161	* __account_scheduler_latency - record an occurred latency
ad0b0fd5 AV	162	* @tsk - the task struct of the task hitting the latency
	163	* @usecs - the duration of the latency in microseconds
	164	* @inter - 1 if the sleep was interruptible, 0 if uninterruptible
	165	*
	166	* This function is the main entry point for recording latency entries
	167	* as called by the scheduler.
	168	*
	169	* This function has a few special cases to deal with normal 'non-latency'
	170	* sleeps: specifically, interruptible sleep longer than 5 msec is skipped
	171	* since this usually is caused by waiting for events via select() and co.
	172	*
	173	* Negative latencies (caused by time going backwards) are also explicitly
	174	* skipped.
	175	*/
9745512c	176	void __sched
ad0b0fd5	177	__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
9745512c AV	178	{
	179	unsigned long flags;
	180	int i, q;
	181	struct latency_record lat;
	182
9745512c AV	183	/* Long interruptible waits are generally user requested... */
	184	if (inter && usecs > 5000)
	185	return;
	186
ad0b0fd5 AV	187	/* Negative sleeps are time going backwards */
	188	/* Zero-time sleeps are non-interesting */
	189	if (usecs <= 0)
	190	return;
	191
9745512c AV	192	memset(&lat, 0, sizeof(lat));
	193	lat.count = 1;
	194	lat.time = usecs;
	195	lat.max = usecs;
f9387721 TG	196
f9387721 TG	197	stack_trace_save_tsk(tsk, lat.backtrace, LT_BACKTRACEDEPTH, 0);
9745512c	198
757455d4	199	raw_spin_lock_irqsave(&latency_lock, flags);
9745512c AV	200
	201	account_global_scheduler_latency(tsk, &lat);
	202
38715258	203	for (i = 0; i < tsk->latency_record_count; i++) {
9745512c AV	204	struct latency_record *mylat;
9745512c AV	205	int same = 1;
19fb518c	206
9745512c	207	mylat = &tsk->latency_record[i];
ad0b0fd5	208	for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
19fb518c DA	209	unsigned long record = lat.backtrace[q];
	210
	211	if (mylat->backtrace[q] != record) {
9745512c	212	same = 0;
9745512c	213	break;
19fb518c DA	214	}
19fb518c DA	215
accddc41 TG	216	/* 0 entry is end of backtrace */
accddc41 TG	217	if (!record)
9745512c AV	218	break;
	219	}
	220	if (same) {
	221	mylat->count++;
	222	mylat->time += lat.time;
	223	if (lat.time > mylat->max)
	224	mylat->max = lat.time;
	225	goto out_unlock;
	226	}
	227	}
	228
38715258 KC	229	/*
	230	* short term hack; if we're > 32 we stop; future we recycle:
	231	*/
	232	if (tsk->latency_record_count >= LT_SAVECOUNT)
	233	goto out_unlock;
	234
9745512c	235	/* Allocated a new one: */
38715258	236	i = tsk->latency_record_count++;
9745512c AV	237	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
	238
	239	out_unlock:
757455d4	240	raw_spin_unlock_irqrestore(&latency_lock, flags);
9745512c AV	241	}
	242
	243	static int lstats_show(struct seq_file m, void v)
	244	{
	245	int i;
	246
	247	seq_puts(m, "Latency Top version : v0.1\n");
	248
	249	for (i = 0; i < MAXLR; i++) {
34e49d4f JP	250	struct latency_record *lr = &latency_record[i];
	251
	252	if (lr->backtrace[0]) {
9745512c	253	int q;
34e49d4f JP	254	seq_printf(m, "%i %lu %lu",
34e49d4f JP	255	lr->count, lr->time, lr->max);
9745512c	256	for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
34e49d4f	257	unsigned long bt = lr->backtrace[q];
accddc41	258
34e49d4f	259	if (!bt)
9745512c	260	break;
accddc41	261
34e49d4f	262	seq_printf(m, " %ps", (void *)bt);
9745512c	263	}
eaa1809b	264	seq_puts(m, "\n");
9745512c AV	265	}
	266	}
	267	return 0;
	268	}
	269
	270	static ssize_t
	271	lstats_write(struct file file, const char __user buf, size_t count,
	272	loff_t *offs)
	273	{
	274	clear_global_latency_tracing();
	275
	276	return count;
	277	}
	278
	279	static int lstats_open(struct inode inode, struct file filp)
	280	{
	281	return single_open(filp, lstats_show, NULL);
	282	}
	283
97a32539 AD	284	static const struct proc_ops lstats_proc_ops = {
	285	.proc_open = lstats_open,
	286	.proc_read = seq_read,
	287	.proc_write = lstats_write,
	288	.proc_lseek = seq_lseek,
	289	.proc_release = single_release,
9745512c AV	290	};
	291
	292	static int __init init_lstats_procfs(void)
	293	{
97a32539	294	proc_create("latency_stats", 0644, NULL, &lstats_proc_ops);
988f11e0	295	#ifdef CONFIG_SYSCTL
	296	register_sysctl_init("kernel", latencytop_sysctl);
	297	#endif
9745512c AV	298	return 0;
9745512c AV	299	}
ad0b0fd5	300	device_initcall(init_lstats_procfs);